Research Scientist / Research Engineer - Foundations, Search

About the Role

We’re looking for a researcher focused on our search efforts. You’ll work with a a team of world-class research scientists and engineers developing foundational technology that enables models to retrieve and condition on the right information, at the right time. This includes designing new embedding training objectives, scalable vector store architectures, and dynamic indexing methods.

This work will support retrieval across many OpenAI products and internal research efforts, with opportunities for scientific publication and deep technical impact.

This role is based in San Francisco, CA. We use a hybrid work model of 3 days in the office per week and offer relocation assistance to new employees.

Responsibilities

• Tackle embedding models and retrieval systems optimized for grounding, relevance, and adaptive reasoning.
• Collaborate with a team of researchers and engineers building end-to-end infrastructure for training, evaluating, and integrating embeddings into frontier models.
• Drive innovation in dense, sparse, and hybrid representation techniques, metric learning, and learning-to-retrieve systems.
• Collaborate closely with Pretraining, Inference, and other Research teams to integrate retrieval throughout the model lifecycle
• Contribute to OpenAI’s long-term vision of AI systems with memory and knowledge access capabilities rooted in learned representations.

You Might Thrive in This Role If You Have

• Proven experience leading high-performance teams of researchers or engineers in ML infrastructure or foundational research.
• Deep technical expertise in representation learning, embedding models, or vector retrieval systems.
• Familiarity with transformer-based LLMs and how embedding spaces can interact with language model objectives.
• Research experience in areas such as contrastive learning, supervised or unsupervised embedding learning, or metric learning.
• A track record of building or scaling large machine learning systems, particularly embedding pipelines in production or research contexts.
• A first-principles mindset for challenging assumptions about how retrieval and memory should work for large models.